Despite decades of clinical trial experience, our understanding of how best to achieve durable recovery from major depression remains very limited. The aim of this study is to define treatment-relevant phenotypes of depression in order to aid practicing clinicians in achieving that goal. We hypothesize that a small, clinically practical set of variables including: 1) dimensional measures of mood disorder and common anxiety comorbidities (embodied in assessment instruments emphasizing a spectrum approach to description of clinical phenotypes); and 2) treatment exposure (embodied in population pharmacokinetic measures for pharmacotherapy and treatment specificity for psychotherapy); will be statistically significant moderators of time to stabilizaton of a major depressive episode. We hypothesize that this same set of variables will be statistically significant mediators of time to relapse and residual functional impairment. Finally we hypothesize that, using signal detection analysis methods to examine the relationship of these variables and traditional correlates of treatment outcome such as baseline severity, residual symptoms and Axis It comorbidity, we can develop clinically useful algorithms to guide clinicians in choosing between pharmacotherapy and psychotherapy as an initial treatment strategy. In order to test these hypotheses, we will randomly assign 288 men and women between 18 and 64 years of age who are seeking treatment for a major depressive episode at local community clinics to begin treatment with either interpersonal psychotherapy (IPT) or SSRI (citalopram) pharmacotherapy. Those who stabilize (Hamilton Rating Scale for Depression score < 7 X 3 weeks) will continue in their initial treatment. Those who do not, will have the other treatment added to their regimen. All stabilizing subjects will enter a continuation treatment phase in which the treatment that brought about the remission (SSRI alone, IPT alone, or the combination) will be continued for 6 months. Our interest is in identifying those subgroups of patients that respond best to specific treatments or treatment sequences, achieve full remission of symptoms and return of functioning, and are able to sustain their recovery and improved functioning through an extended well interval. By using a relatively small set of variables that we believe have high potential for outcome prediction to characterize patients and their treatment, we expect to accomplish this goal in the context of the proposed study. Cox proportional hazard survival regression models and random regression models will be used to analyze the association of time to stabilization and time to relapse and degree of functional impairment) with spectrum assessments and treatment exposure variables. Signal detection analysis will be used to determine which combination of spectrum assessment scores and other clinical variables describe the profile of patients likely to stabilize or relapse with each of the treatments.